Evaluating Privacy and Utility in Synthetic EHR Data Generation for Adverse Drug Event Detection

Area of Science:

• Health Informatics
• Machine Learning
• Data Privacy

Background:

• Electronic Health Records (EHR) contain valuable information for detecting Adverse Drug Events (ADEs).
• Generating realistic synthetic EHR data is crucial for research and development while protecting patient privacy.
• The Synthetic Data Vault (SDV) tool offers various models for synthetic data generation.

Purpose of the Study:

• To evaluate the effectiveness of different SDV models in generating synthetic EHR data for ADE detection.
• To assess the utility, fidelity, and privacy of synthetic data generated by GaussianCopula, CTGAN, and TVAE models.
• To determine the optimal synthetic data generation strategy based on dataset characteristics and application requirements.

Main Methods:

• Utilized three SDV models: GaussianCopula, Conditional Tabular Generative Adversarial Network (CTGAN), and Tabular Variational Autoencoder (TVAE).
• Employed a structured Swedish EHR dataset for experiments.
• Evaluated synthetic data using SynthEval metrics and a 'train-on-synthetic, test-on-real' (TSTR) approach with Random Forest classifiers.

Main Results:

• TVAE performance was dependent on dataset size and class balance, improving with larger datasets.
• GaussianCopula demonstrated stable utility and enhanced privacy but lower fidelity.
• CTGAN produced realistic data but showed variable performance in TSTR evaluations.

Conclusions:

• The choice of synthetic data generation model significantly influences ADE detection performance.
• Model selection should consider specific healthcare application needs and the characteristics of the available dataset.
• Balancing data utility, fidelity, and privacy is essential when generating synthetic EHR data.